Re: [PATCH v2 3/3] vhost: add device op to offload the interrupt kick

[Eelco Chaudron <echaudro@redhat.com>]

On 16 May 2023, at 12:12, David Marchand wrote:

> On Tue, May 16, 2023 at 10:53 AM Eelco Chaudron <echaudro@redhat.com> wrote:
>> On 10 May 2023, at 13:44, David Marchand wrote:
>
> [snip]
>
>>>> @@ -846,6 +848,14 @@ vhost_user_socket_mem_free(struct vhost_user_socket *vsocket)
>>>>                 vsocket->path = NULL;
>>>>         }
>>>>
>>>> +       if (vsocket && vsocket->alloc_notify_ops) {
>>>> +#pragma GCC diagnostic push
>>>> +#pragma GCC diagnostic ignored "-Wcast-qual"
>>>> +               free((struct rte_vhost_device_ops *)vsocket->notify_ops);
>>>> +#pragma GCC diagnostic pop
>>>> +               vsocket->notify_ops = NULL;
>>>> +       }
>>>
>>> Rather than select the behavior based on a boolean (and here force the
>>> compiler to close its eyes), I would instead add a non const pointer
>>> to ops (let's say alloc_notify_ops) in vhost_user_socket.
>>> The code can then unconditionnally call free(vsocket->alloc_notify_ops);
>>
>> Good idea, I will make the change in v3.
>
> Feel free to use a better name for this field :-).
>
>>
>>>> +
>>>>         if (vsocket) {
>>>>                 free(vsocket);
>>>>                 vsocket = NULL;
>
> [snip]
>
>>>> +       /*
>>>> +        * Although the ops structure is a const structure, we do need to
>>>> +        * override the guest_notify operation. This is because with the
>>>> +        * previous APIs it was "reserved" and if any garbage value was passed,
>>>> +        * it could crash the application.
>>>> +        */
>>>> +       if (ops && !ops->guest_notify) {
>>>
>>> Hum, as described in the comment above, I don't think we should look
>>> at ops->guest_notify value at all.
>>> Checking ops != NULL should be enough.
>>
>> Not sure I get you here. If the guest_notify passed by the user is NULL, it means the previously ‘reserved[1]’ field is NULL, so we do not need to use a new structure.
>>
>> I guess your comment would be true if we would introduce a new field in the data structure, not replacing a reserved one.
>
> Hum, I don't understand my comment either o_O'.
> Too many days off... or maybe my evil twin took over the keyboard.
>
>
>>
>>>> +               struct rte_vhost_device_ops *new_ops;
>>>> +
>>>> +               new_ops = malloc(sizeof(*new_ops));
>>>
>>> Strictly speaking, we lose the numa affinity of "ops" by calling malloc.
>>> I am unclear of the impact though.
>>
>> Don’t think there is a portable API that we can use to determine the NUMA for the ops memory and then allocate this on the same numa?
>>
>> Any thoughts or ideas on how to solve this? I hope most people will memset() the ops structure and the reserved[1] part is zero, but it might be a problem in the future if more extensions get added.
>
> Determinining current numa is doable, via 'ops'
> get_mempolicy(MPOL_F_NODE | MPOL_F_ADDR), like what is done for vq in
> numa_realloc().
> The problem is how to allocate on this numa with the libc allocator
> for which I have no idea...
> We could go with the dpdk allocator (again, like numa_realloc()).
>
>
> In practice, the passed ops will be probably from a const variable in
> the program .data section (for which I think fields are set to 0
> unless explicitly initialised), or a memset() will be called for a
> dynamic allocation from good citizens.
> So we can probably live with the current proposal.
> Plus, this is only for one release, since in 23.11 with the ABI bump,
> we will drop this compat code.
>
> Maxime, Chenbo, what do you think?

Wait for their response, but for now I assume we can just keep the numa unaware malloc().

>
> [snip]
>
>>>
>>> But putting indentation aside, is this change equivalent?
>>> -               if ((vhost_need_event(vhost_used_event(vq), new, old) &&
>>> -                                       (vq->callfd >= 0)) ||
>>> -                               unlikely(!signalled_used_valid)) {
>>> +               if ((vhost_need_event(vhost_used_event(vq), new, old) ||
>>> +                               unlikely(!signalled_used_valid)) &&
>>> +                               vq->callfd >= 0) {
>>
>> They are not equal, but in the past eventfd_write() should also not have been called with callfd < 0, guess this was an existing bug ;)
>
> I think this should be a separate fix.

ACK, will add a separate patch in this series to fix it.

>
>>
>>>> +                       vhost_vring_inject_irq(dev, vq);
>
>
> -- 
> David Marchand